Disney found out Sora was shutting down one hour before the rest of the world.
They had committed $1 billion to the partnership.
That's not a communication mishap. That's a signal about how deeply the product was actually integrated — and how much weight the partnership really carried beyond the press release.
Sora launched in September 2025. It went viral immediately, as you'd expect. OpenAI, AI video generation, millions of users — every checkbox was ticked. Six months later it was gone. User count had dropped from a million to under 500,000. It was burning roughly $1 million a day. Deepfake scandals kept erupting: Michael Jackson, Martin Luther King, Mister Rogers — all appearing in AI-generated scenarios their estates and representatives hadn't consented to.
OpenAI pulled the plug.
The model worked. The technology was real. Something else failed.
Unit economics that never closed.
This is the least-discussed reason and probably the most fundamental. A million dollars a day in infrastructure costs — not overhead, actual video generation costs. And the revenue per user didn't come close to covering it. Worse, users weren't staying. Growth was running backwards: from a million toward 500,000. That's not a retention problem to be optimized. That's a signal that the product hadn't found a stable use case yet.
A product manager's early question should have been: At what user count, at what usage intensity, does this unit economics actually close? If the answer was genuinely uncertain, then Sora wasn't a consumer product — it was infrastructure research wearing a consumer product's clothes. And infrastructure research doesn't get a splashy public launch.
Trust infrastructure should precede content policy, not follow it.
The deepfake problem was foreseeable. Almost certainly foreseen. The issue wasn't awareness — it was sequencing: ship first, figure out content safety after.
For some products, that works. You move fast, you iterate, users are patient, the stakes are low. AI video generation isn't that product. Video of a public figure doing something they didn't do isn't a bug you can patch. Once it's out, it's out. The actor's union complaints, the estate lawsuits, the public backlash — all of this arrived after launch. Which is too late.
The PM decision that mattered wasn't what the model could do. It was: Can we handle what users will do with it? That's a different question, and it needs a different kind of rigor — not just technical testing, but adversarial scenario planning before the first user ever logs in.
What an enterprise partnership actually means.
Back to Disney. A billion-dollar partner found out their investment was being dissolved with an hour's notice.
That tells you the partnership existed on paper more than in practice. Enterprise relationships in AI aren't just contracts — they're bets on a product's future. And when someone bets a billion dollars on your product's future, they should be in the room for critical product decisions. Not as a veto, but as a stakeholder who understands where things are heading.
If Disney's product team had real visibility into Sora's user trajectory and burn rate over those six months, the shutdown wouldn't have been a surprise. The surprise itself is the symptom.
I keep thinking about what the PM conversations inside OpenAI looked like in months three and four. When user numbers started declining. When the $1M/day burn wasn't showing signs of coming down. When the deepfake incidents kept recurring despite the crackdowns.
Did the question "should we continue this product?" come up early enough to matter? Or did momentum — the public launch, the Disney deal, the press coverage — make it difficult to say "this isn't working" out loud?
That's a real PM failure mode. Not incompetence. Just the normal human difficulty of calling a thing that has already been celebrated a mistake.
OpenAI has built some remarkable things. Sora's failure isn't a technology story — the model worked. It's a story about product decisions made at the wrong time, in the wrong sequence, with the wrong questions asked too late.
The strongest technology is not immune to bad product judgment. That's precisely why Sora is worth remembering.